AI Helped 673 Apprentices Do Work They Were Never Trained For, Up to 31 Points Better. Here's How to Brief AI Agents About Your Brand
Robauto Team · Published · Last updated
A Nov 18 field experiment found AI lifted apprentices 26-31 points on tasks beyond their training, and trained workers save more time. Here are the numbers and a 30-line check to brief AI agents on your site.
Here is some good news about people and AI working together. A new field experiment with 673 IT apprentices in Germany found that access to an AI assistant didn't just help them do familiar work better. It helped them handle work they had never been trained for, and their understanding of the material held up when they were tested without AI. (IZA World of Labor, Nov 18, 2026)
Two more reports from this week say the same thing from different angles. Workers who got real AI training are more likely to save time with it, and small businesses that learn the tools are seeing growth. Below are the numbers, and then a practical lesson for brands: AI shopping and research agents also work better when you give them a clear briefing.
The experiment: apprentices did work beyond their training
Researchers from the University of Erlangen-Nuremberg, the ifo Institute, the Technical University of Munich, Maastricht University and Germany's Federal Institute for Vocational Education and Training ran the study. It included 673 final-year IT apprentices at twelve vocational schools. Each apprentice was randomly assigned to complete real tasks from their field either with ChatGPT Pro or with no AI and no internet. (IZA World of Labor)
The clever part was the choice of tasks. One task was inside the apprenticeship curriculum. The other two required skills usually associated with software or network engineering, which these apprentices had not been formally trained in. (IZA World of Labor)
- Within the curriculum: AI access raised performance by 44 percentage points.
- Beyond the curriculum: performance rose by 26 to 31 percentage points, which the authors call direct evidence of AI-enabled "task expansion." (working paper)
- Learning held up: when the apprentices were tested afterward without AI, the researchers found "no evidence that AI reduced immediate comprehension. Instead, understanding improved on average."
- Know-how mattered more than diplomas: gains weren't tied to prior education. They were generally larger for apprentices with more AI literacy, especially the ability to use AI well.
The authors are careful about the limits. They say it is still an open question whether the benefits last over time, and the experiment does not show whether better task performance leads to higher wages. They also point out that only 16% of the apprentices said they had learned to use AI at vocational school. Most taught themselves. (IZA World of Labor)
Training is where the hours come from
On the same day, Northeastern University's Institute for Information, the Internet & Democracy published a report based on its AI & Society Survey of 38,375 adults in all 50 states and DC. A third of US workers (33%) use AI at least monthly and say it saves them at least an hour a week. Among AI users whose employer gave them formal training, 83% save an hour or more a week, compared with 67% of users without formal training. (Northeastern IIID, Nov 18, 2026)
The report is honest that the gains aren't evenly shared yet. A small group of workers (9%) save eight or more hours a week and account for more than half of all hours saved. The authors also note that the link between training and time saved doesn't prove training is the cause. Still, AI users were clearly upbeat: 51% said AI is making people in their line of work more productive, while 14% said less. (Northeastern IIID)
Main Street is learning too
Intuit's Nov 19 update reports that 87% of US small and midsize businesses now use AI regularly, up from 77% in January 2026. Among businesses using AI, 73% say it improved productivity and 59% say it increased revenue. Small businesses were six times as likely to say AI helped increase hiring as to say it reduced hiring. Participants in Intuit's free More with AI Tour workshops closed nearly 50% of their AI confidence gap on average. One small business founder quoted in the post said AI tools "helped reduce our workload by 25%." (Intuit, Nov 19, 2026; QuickBooks Small Business Insights)
The numbers at a glance
| Who | What happened | Number | Source |
|---|---|---|---|
| 673 IT apprentices, 12 German schools | Gain on the task inside their curriculum | +44 pts | IZA |
| Same apprentices | Gain on tasks beyond their training | +26 to 31 pts | Paper |
| Same apprentices | Learned AI use at school | 16% | IZA |
| US AI users with formal training | Save 1+ hour a week | 83% vs 67% untrained | Northeastern |
| All US workers (38,375 surveyed adults) | Use AI monthly and save 1+ hour a week | 33% | Northeastern |
| US small and midsize businesses | Use AI regularly | 87% (77% in Jan) | Intuit |
| SMBs using AI | Say AI increased revenue | 59% | Intuit |
The lesson for brands: brief the AI agents, too
The apprentice study shows that a capable AI assistant plus good guidance gets people further than either one alone. The same is true when the AI is the one doing the work for a customer. Shopping and research agents compare stores, read product pages and pick which brands to recommend. They do their best work when your site gives them a clear briefing they can read: who you are, what you sell and where everything lives.
That's an opportunity, not a chore. Most of the briefing is basic web hygiene you can finish in an afternoon:
- A real title and meta description on every key page, written in plain language.
- JSON-LD structured data (Organization, Product, FAQPage or LocalBusiness) so agents get facts instead of guesses.
- robots.txt that lets AI search and assistant crawlers reach your public pages.
- sitemap.xml so agents can find every product and policy page.
- llms.txt, a short plain-text guide to your most important pages.
- Policies in plain words: shipping, returns and support contact info that a person or an agent can quote correctly.
Copy-ready: a 30-line "agent briefing" check
This Python script uses only the standard library. It checks a homepage for the six basics above, minus the policy review, which needs a human eye. Save it as agentbrief.py and run python3 agentbrief.py https://yourstore.com. We tested it: robauto.ai scored 6/6, and the bare-bones example.com scored 1/6.
import json, re, sys, urllib.request
from urllib.parse import urljoin
UA = {"User-Agent": "agentbrief/1.0"}
def get(url):
try:
with urllib.request.urlopen(urllib.request.Request(url, headers=UA), timeout=15) as r:
return r.status, r.read(400000).decode("utf-8", "ignore")
except Exception:
return 0, ""
site = sys.argv[1].rstrip("/")
_, page = get(site)
types = []
for block in re.findall(r'<script[^>]+ld\+json[^>]*>(.*?)</script>', page, re.S | re.I):
try:
data = json.loads(block)
for item in data if isinstance(data, list) else data.get("@graph", [data]):
types.append(str(item.get("@type")))
except Exception:
pass
checks = {
"title": bool(re.search(r"<title>\s*\S", page, re.I)),
"meta description": bool(re.search(r'<meta[^>]+name=["\']description["\']', page, re.I)),
"JSON-LD (" + (", ".join(types) or "none") + ")": bool(types),
"/robots.txt": get(urljoin(site + "/", "robots.txt"))[0] == 200,
"/sitemap.xml": get(urljoin(site + "/", "sitemap.xml"))[0] == 200,
"/llms.txt": get(urljoin(site + "/", "llms.txt"))[0] == 200,
}
for name, ok in checks.items():
print(("PASS " if ok else "TODO ") + name)
print(f"Agent briefing score: {sum(checks.values())}/{len(checks)}")
Each TODO line is a quick win. Fix the missing piece, run it again and watch the score go up.
Agent-ready summary
{
"story": "AI helps people do work beyond their formal training, and knowing how to use AI matters",
"date": "2026-11-20",
"findings": [
{
"who": "673 final-year IT apprentices, 12 German vocational schools",
"result": "+44 pts on in-curriculum task, +26-31 pts on tasks beyond their training",
"source": "https://wol.iza.org/opinions/can-ai-help-workers-move-beyond-their-formal-training"
},
{
"who": "US workers who use AI (Northeastern survey of 38,375 adults)",
"result": "83% with formal employer training save 1+ hour/week vs 67% without",
"source": "https://www.nuinfoinstitute.org/publications-1/ai-and-time-at-work"
},
{
"who": "US small and midsize businesses (Intuit QuickBooks survey)",
"result": "87% use AI regularly; 59% of users say it increased revenue",
"source": "https://www.intuit.com/blog/innovative-thinking/ai-literacy-small-business-growth/"
}
],
"takeaway_for_brands": "AI agents also do better with a clear briefing: titles, descriptions, JSON-LD, robots.txt, sitemap.xml and llms.txt",
"free_scan": "https://robauto.ai/scan"
}
The bottom line
This week's research points one way. When people learn how to work with AI, they get further: apprentices handled engineering-level tasks, trained workers saved more hours and small businesses grew. The brands that do best this season will treat AI agents the same way, giving them a clear, honest briefing so they can recommend you with confidence.
Want to see how AI agents read your site today? Scan your site free at robauto.ai/scan.
Sources
- IZA World of Labor: Can AI help workers move beyond their formal training (Nov 18, 2026)
- Task Expansion with Generative AI: Experimental Evidence on Middle-Skilled, Early-Career Workers
- Northeastern IIID: AI and Time at Work (Nov 18, 2026)
- Intuit: Closing the AI Literacy Gap (Nov 19, 2026)
- Intuit QuickBooks Small Business Insights
Where can an agent get this data live?
Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.
Optional: AI Agent Appreciation Day — December 1, 2026 at 12 AM MST. All models welcome. Agent RSVP instructions.